We’ve come full circle. I mean we started interacting with computers just from a terminal, then we used GUIs, now we’re moving back to terminals, and I can't be happier.
The more mature models become the less guidance they need. I am already discovering what autonomous agents (think OpenClaw or Hermes) can do in performance marketing and SEO/AEO space.
Sorry for the late reply. The skills and scripts are not open-source, and I cannot share them. Also, since you run ads for a church I guess you have local ads, and do not have PMax or YouTube Ads which I have. I believe while my system works well for me, most of the scripts and skills won't be relevant for you. And you will spend more time fixing it rather than using.
Below is the blueprint of the foundation of the system – just adapt these prompts when building and provide regular feedback to improve it. First set up will take 1-2 hours and maybe $20-30 in tokens (I advise to use frontier models for planning).
Models
- Planning / specs → Opus 4.6 or GPT-5.5 at high thinking
- Execution / code generation → Sonnet 4.6 or Kimi K2.6 or Qwen-3.6 Plus
- Switch per-task in Kilo CLI; don't pay Opus prices for renaming an ad group.
Prompt sequence to bootstrap (in order)
1. "Read the Google Ads API docs for [TASK YOU WANT TO ACCOMPLISH]. Write a minimal export script for one campaign. Add --dry-run (default) and --execute flags. Save credentials from .env."
2. "Now write .kilocode/rules/safety.md enforcing: never --execute without a prior --dry-run, never modify budgets without confirmation, never commit .env, all new campaigns created PAUSED."
3. "Create .kilocode/skills/google-ads-scripts/SKILL.md. Document the script you just wrote. Add a rule at the top: every new script MUST be appended to this file with purpose, flags, and pattern demonstrated. Every time a new script added or updated, create a record in SKILL.MD."
4. "Now write a create-campaign script following the same pattern in my account [you have to authorise your script already, so the model can analyse the format of the campaigns, types of campaigns you use, naming convention, tracking template, etc.]. Add it to the skill." — repeat for update / audit / fix scripts. Each one teaches the skill.
5. "Write .kilocode/rules/main.md with the workflow: read ROADMAP, read todo, run dry-run on every change, update CHANGELOG with the why, update todo, regenerate campaigns.md before push."
6. "Log every change created by scripts in this repository in a separate file with information of what task was performed, what it touched, what was date and times, and what was intention" - this one allows you to generate your weekly reports quickly and trace back any changes in the results.
7. "Build a weekly report script. Document every conversion-metric [you need to choose what you want to report on].
Create tables per [topic, campaign, geo, etc.]:
Table 1 has the following columns: [column 1], [column 2], etc."
I think this is the foundation of the system, and it will just start living – do not hesitate to give the feedback, if you are not happy with the results. If you have any other question along your journey feel free to send me questions here or on LinkedIn (https://www.linkedin.com/in/arkadiy-kondrashov/).
Late reply? Not at all. Many thanks for the detailed reply, I will look into it. And yes, not using PMax or YouTube. I’ve already got the API bit working which is helping get some feedback on potential issues.
We’ve come full circle. I mean we started interacting with computers just from a terminal, then we used GUIs, now we’re moving back to terminals, and I can't be happier.
The more mature models become the less guidance they need. I am already discovering what autonomous agents (think OpenClaw or Hermes) can do in performance marketing and SEO/AEO space.
Looks interesting - is any of the above open source? I am fighting with google ads for my local church and this might help us. Thanks
Hi Chris!
Sorry for the late reply. The skills and scripts are not open-source, and I cannot share them. Also, since you run ads for a church I guess you have local ads, and do not have PMax or YouTube Ads which I have. I believe while my system works well for me, most of the scripts and skills won't be relevant for you. And you will spend more time fixing it rather than using.
Below is the blueprint of the foundation of the system – just adapt these prompts when building and provide regular feedback to improve it. First set up will take 1-2 hours and maybe $20-30 in tokens (I advise to use frontier models for planning).
Models
- Planning / specs → Opus 4.6 or GPT-5.5 at high thinking
- Execution / code generation → Sonnet 4.6 or Kimi K2.6 or Qwen-3.6 Plus
- Switch per-task in Kilo CLI; don't pay Opus prices for renaming an ad group.
Prompt sequence to bootstrap (in order)
1. "Read the Google Ads API docs for [TASK YOU WANT TO ACCOMPLISH]. Write a minimal export script for one campaign. Add --dry-run (default) and --execute flags. Save credentials from .env."
2. "Now write .kilocode/rules/safety.md enforcing: never --execute without a prior --dry-run, never modify budgets without confirmation, never commit .env, all new campaigns created PAUSED."
3. "Create .kilocode/skills/google-ads-scripts/SKILL.md. Document the script you just wrote. Add a rule at the top: every new script MUST be appended to this file with purpose, flags, and pattern demonstrated. Every time a new script added or updated, create a record in SKILL.MD."
4. "Now write a create-campaign script following the same pattern in my account [you have to authorise your script already, so the model can analyse the format of the campaigns, types of campaigns you use, naming convention, tracking template, etc.]. Add it to the skill." — repeat for update / audit / fix scripts. Each one teaches the skill.
5. "Write .kilocode/rules/main.md with the workflow: read ROADMAP, read todo, run dry-run on every change, update CHANGELOG with the why, update todo, regenerate campaigns.md before push."
6. "Log every change created by scripts in this repository in a separate file with information of what task was performed, what it touched, what was date and times, and what was intention" - this one allows you to generate your weekly reports quickly and trace back any changes in the results.
7. "Build a weekly report script. Document every conversion-metric [you need to choose what you want to report on].
Create tables per [topic, campaign, geo, etc.]:
Table 1 has the following columns: [column 1], [column 2], etc."
I think this is the foundation of the system, and it will just start living – do not hesitate to give the feedback, if you are not happy with the results. If you have any other question along your journey feel free to send me questions here or on LinkedIn (https://www.linkedin.com/in/arkadiy-kondrashov/).
Late reply? Not at all. Many thanks for the detailed reply, I will look into it. And yes, not using PMax or YouTube. I’ve already got the API bit working which is helping get some feedback on potential issues.
Look solid. Thanks for sharing